Pingtou Ge's one-stop chip design platform "Swordless" is out

In the first generation of chip design, the circuit had to be redesigned every time for different products. As the iteration of chips became faster, IP-based design methods began to emerge after 2000, which is called the 2.0 era of chip design.

As the era of the Internet of Everything approaches, the requirements for chip cost and performance are getting higher and higher. As a result, people are beginning to look forward to a new chip design method with higher efficiency and lower cost. Pingtou Ge has keenly sensed this and brought the latest one-stop chip design platform - "Wujian", which is aimed at the era of chip design 3.0!

Why are costs and design cycles cut in half?

In July this year, Pingtou Ge released the industry's most powerful RISC-V processor - Xuantie 910. Its name comes from the "Xuantie Heavy Sword" used by Dugu Qiubai before he was forty years old in Jin Yong's novel.

The "Swordless" chip platform released today (29) is also derived from the state of being better without a sword than with a sword after Dugu Qiubai turned 40. Just like the Swordless platform of Pingtou Ge does not have a chip, it can help chip design companies to "forge swords" efficiently.

According to Jiwei.com, the Wujian SoC chip platform can shorten the design cycle of AIoT chips by more than 50% and reduce costs by more than 50%. How is this done?

Meng Jianyi, a researcher at Pingtou Ge Semiconductor, pointed out that the Wujian platform is composed of modules such as SoC architecture, processor, various IPs, operating system, software driver and development tools, and is a basic common technology platform for system chip development.

In terms of cost reduction, Wujian's platform-based design method that combines software and hardware enables IP to be quickly connected to the system, reducing the price of IP and the manpower investment required for R&D.

The main reason for shortening the design cycle is that the Wujian platform can reduce the repetitive input in the chip design process. In addition, according to Meng Jianyi, more than 50% of the design verification work can be eliminated, because if all the analog IP and foundry processes of the platform have been verified, the trial production (MPW) stage of more than 3 months can be skipped and mass production can be directly entered, which can shorten the time from chip design to mass production to less than 9 months.

The first move of “no sword” swordsmanship——Visual AI platform

Based on the Wujian platform, Pingtou Ge also released the Wujian visual AI platform today based on the high-performance Xuantie full series CPU. This platform includes 128-bit single instruction multiple data SIMD vector extension technology, which provides sufficient computing power on the processor side, and the main frequency support ranges from 1-2GHz.

As the first move of Pingtou Ge's "Swordless" swordsmanship, the biggest feature of this visual AI platform is that it can support third-party AI accelerators. It is understood that it can support neural network accelerators with computing power ranging from 0.5-16Tops. It is an open architecture that can support a maximum storage bandwidth of 400Gbps. The bandwidth can be matched according to different computing power, and will support HBM and other higher bandwidth storage in the future.

In addition, the Wujian Vision AI platform can also support the expansion of multiple small chips into chips with stronger computing power, which greatly improves the adaptability of the chip. Not only that, the platform is also open to IP companies, allowing the world's most competitive IP products to conduct prototype tape-out verification with the Wujian platform, providing silicon-verified mass-producible chip-level overall solutions.

The first form of swordless has three major characteristics in terms of basic attributes:

1. Ultra-low power multi-core architecture: The platform adopts a large and small core architecture. The large core supports the neural network accelerator to achieve high-performance computing, and the small core manages the on-chip power network and peripheral interfaces in the manner of an MCU. This architecture can effectively control the startup time of the entire system to less than 200ms.

2. Security architecture: The platform provides hardware-supported TEE security technology, which can effectively isolate secure information and prevent applications from directly accessing secure resources. This technology has passed the compatibility certification of Global Platform.

3. Online performance analysis: In the past, chips could not provide real-time data in online situations during design and mass production, but the Wujian AI platform can analyze computing power bandwidth in real time, allowing users to capture customized events, and also supports data export from different interfaces such as DDR, USB, PCIE, and JTAG. This makes the development and performance optimization of AI chips more user-friendly.

What is the best way to tap into the potential capabilities of chips? Pingtou Ge believes that it is a combination of software and hardware. So the company introduced the neural network model compression technology provided by the Aurora software team, which is a structured compression method based on automatic sparsity perception based on reinforcement learning. By combining with the hardware of the Wujian platform, it can provide a software-coordinated sparsification algorithm, thereby compressing the space of the neural network model by an average of more than 60%, and improving the performance by an average of more than 60%.

In addition to the combination of software and hardware, Pingtou Ge has also prepared a set of tools - a full-stack AI development kit, which supports AI acceleration engines that can be configured in multiple scenarios through a unified AI software development framework, and realizes the ability to deploy applications with one click and graphical computing power analysis. The cooperation practice with third-party companies shows that the cost of basic software development for AI chips based on the full-stack AI development kit has been reduced by 60%, and the user application development cycle has also been reduced by 50%.

Create a "Flathead Brother Model" that leads the times

Meng Jianyi said that the Wujian SoC platform can be obtained through open licensing, and its goal is to be integrated. Pingtou Ge hopes to create a one-stop chip design model - the "Pingtou Ge model" to lead the chip design 3.0 era.

The three main features of the "Pingtou Ge Model" are: full stack, open and integrated. This is a full-stack integration model from chips to applications, which can achieve the deep integration of core software and hardware technologies such as processors, algorithms, operating systems, etc., breaking the limitations of the traditional general-purpose chip era's IP licensing commercial model, which is high cost, difficult to use, and long cycle. It provides companies with full-stack technical capabilities for chip design and opens this capability to the whole society.

Regarding the future strategic layout, Pingtou Ge pointed out that it will develop infrastructure and peripheral technologies for industrial, automotive, security, access and other application fields, including CPU, SoC platform and field-related security, reliability and other technologies.

Meng Jianyi emphasized that Pingtou Ge's positioning is to build the infrastructure of the chip industry. In the future, Alibaba will continue to invest heavily in building good technology and building an application ecosystem. In addition to processors, Pingtou Ge will continue to develop operating systems, algorithms that integrate software and hardware, and core IP.

Finally, Meng Jianyi also revealed that the company's first self-developed NPU chip in the cloud will also be released this year. Its performance will be at the leading level in the industry and will be mainly used in Alibaba Cloud data centers.